AI Watch, Friday 4 September 2026OpenAI ships GPT-6 Astra as three assistants go dark

Nvidia confirmed on Thursday 3 September its acquisition of Hugging Face, the platform where open AI models are exchanged (models you can download and run yourself), for $12.93 billion. Founded in 2016 by a French team, it hosts more than 3 million models and serves more than 18 million developers; its co-founder Clément Delangue says he is aiming for 100 million users within two years and vows that openness to the whole ecosystem will be preserved. The deal illustrates Europe's difficulty in keeping its champions: the reference hub for open AI, born in France, passes under an American flag, that of the chipmaker already selling the chips these models run on. For a graphics processor designer, taking hold of the storefront where developers pick their models locks the chain a little tighter, from silicon to use.

Faced with the flood of models on Hugging Face, the CNIL (France's data protection authority) is offering a tool to trace where each one comes from: what data it was trained on, which model it derives from, under which licence it circulates, Next.ink reports. The stakes are concrete for a French organisation that wants to reuse an open model without inheriting an upstream rights or personal-data problem. Here the regulator positions itself as practical support rather than enforcer, at the very moment the open ecosystem changes hands.

Against the market, the free office suite LibreOffice released its 26.8 version with no generative AI features at all, a deliberate choice in the name of data sovereignty, ZDNet reports. The update passed one million downloads in a week. The argument lands with IT departments worried about data leaking to external AI services: guaranteeing that no text goes to a third-party server becomes a selling point.

On Thursday 3 September, the three leading AI assistants suffered major disruptions minutes apart, The Verge and Wired report. Anthropic (Claude) wobbled first in mid-afternoon, followed by ChatGPT and Codex at OpenAI, which returned errors from 11 a.m. in New York (4:30 p.m. in Paris), then Grok at xAI. The services came back after a few hours, but none of the companies explained the cause, and nothing points to a common incident, Wired notes. The simultaneity revives a rarely asked question: these tools, now part of daily work, offer no availability guarantee comparable to that of a critical service.

OpenAI terminated its partnership with Cursor, the maker of an AI-assisted development environment (software that helps write code), after Cursor was bought by SpaceX, Elon Musk's company, Wired reports. The partnership brought OpenAI more than a billion dollars a year, yet it chose to walk away rather than enrich a rival that had become the owner of its customer. The direct consequence for French developers who were using Cursor with OpenAI's models: they must urgently rethink their toolchain and switch to other models, ZDNet notes. The rivalry between Sam Altman and Elon Musk is now paid for in ten-figure commercial breakups.

Google presented WeatherNext 3, an AI weather forecasting model that, according to the company, markedly improves the forecasting of rain and snowfall and produces a global picture at an unprecedented resolution, The Verge and TechCrunch report. The forecasts will feed the weather shown in Google Search, in Google Maps and in the Gemini assistant. Deep-learning weather is gaining ground on classic physics models, which are slower and more compute-hungry. The example shows AI applied to a mass-market service, far from chatbot demos, with a measurable benefit for anyone checking a forecast.

For its new Muse Spark model, meant to run coding agents and other tasks, Meta offers an average discount of about 95% to users who agree to "contribute" to the development of future models by sharing their queries and the responses they get, TechCrunch reports. The deal is explicit: the price of use collapses for anyone who gives up their interaction data. For a company, the trade-off becomes concrete and quantified: saving on the AI bill by feeding a rival's model with its own use cases, some of them sensitive. The transparency of the offer has the merit of making visible an exchange that, elsewhere, stays tacit.

Enflame Technology, one of China's leading AI chip designers, set off a frenzy over its listing on the Shanghai stock exchange, worth 6.12 billion yuan (about $911 million), the South China Morning Post reports. The retail tranche was oversubscribed 4,073 times: nearly 7 million individual investors placed orders for 42.1 billion shares. Backed by Tencent, Enflame presents itself as a local alternative to Nvidia, whose most advanced chips remain hit by American export restrictions. The enthusiasm speaks to the Chinese market's appetite for home-grown silicon champions, as Beijing pushes to replace American technology. The contrast with Europe is stark: the same day, a champion of open AI there passed under American control.

The Beijing start-up Moonshot AI, behind the Kimi K3 model known for its performance, has filed confidentially for a listing in Hong Kong, according to two sources cited by the South China Morning Post. The company is reportedly aiming for a listing as early as the first quarter of 2027 to fund its rapid expansion. Kimi is gaining ground with developers: its coding assistant, Kimi Code, appeals with a promise of far lower cost than Anthropic's Claude Code, the Journal du Net notes. Chinese models are now aiming for Western developers' wallets on price.